Speech recognition, which includes both speaker independent speech recognition and speaker dependent speech recognition, is used for a wide variety of applications.
Speech recognition normally involves the use of speech recognition models or templates that have been trained using speech samples provided by one or more individuals. Commonly used speech recognition models include Hidden Markov Models (HMMS). An example of a common template is a dynamic time warping (DTW) template. In the context of the present application “speech recognition model” is intended to encompass both speech recognition models as well as templates which are used for speech recognition purposes.
As part of a speech recognition operation, speech input is normally digitized and then processed. The processing normally involves extracting feature information, e.g., energy and/timing information, from the digitized signal. The extracted feature information normally takes the form of one or more feature vectors. The extracted feature vectors are then compared to one or more speech recognition models in an attempt to recognize words, phrases or sounds.
In speech recognition systems, various actions, ge.g., dialing a telephone number, entering information into a form, etc., are often performed in response to the results of the speech recognition operation.
Before speech recognition operations can be performed, one or more speech recognition models need to be trained. Speech recognition models can be either speaker dependent or speaker independent. Speaker dependent (SD) speech recognition models are normally trained using speech from a single individual and are designed so that they should accurately recognize the speech of the individual who provided the training speech but not necessarily other individuals. Speaker independent (SI) speech recognition models are normally generated from speech provided from numerous individuals or from text. The generated speaker independent speech recognition models often represent composite models which take into consideration variations between different speakers, e.g., due to differing pronunciations of the same word. Speaker independent speech recognition models are designed to accurately identify speech from a wide range of individuals including individuals who did not provide speech samples for training purposes.
In general, model training involves one or more individuals speaking a word or phrase, converting the speech into digital signal data, and then processing the digital signal data to generate a speech recognition model. Model training frequently involves an iterative process of computing a speech recognition model, scoring the model, and then using the results of the scoring operation to further improve and retrain the speech recognition model.
Speech recognition model training processes can be very computationally complex. This is true particularly in the case of SI models where audio data from numerous speakers is normally processed to generate each model. For this reason, speech recognition models are often generated using a relatively powerful computer systems.
Individual speech recognition models can take up a considerable amount of storage space. For this reason, it is often impractical to store speech recognition models corresponding to large numbers of words or phrases, e.g., the names of all the people in a mid-sized company, or large dictionary in a portable device or speech recognizer where storage space, e.g., memory, is limited.
In addition to limits in storage capacity, portable devices are often equipped with limited processing power. Speech recognition, like the model training process, can be a relatively computationally complex process and can there for be time consuming given limited processing resources. Since most users of a speech processing system expect a prompt response from the system, to satisfy user demands speech processing often needs to be performed in real or near real time. As the number of potential words which may be recognized increases, so does the amount of processing required to perform a speech recognition operation. Thus, devices with limited processing power which may be able to perform a speech recognition operation involving recognizing, e.g., 20 possible names in near real time, may not be fast enough to perform a recognition operation in near real time where the number of names is increased to 100 possible names.
In the case of voice dialing and other applications where the recognition results need to be generated in near real time, e.g., with relatively little delay, the limited processing power of portable devices often limits the size of the vocabulary which can be considered as possible recognition outcomes.
In addition to the above implementation problems, implementers of speech recognition systems are often confronted with logistical problems associated with collecting speech samples to be used for model training purposes. This is particularly a problem in the case of speaker independent speech recognition models where the robustness of the models are often a function of the number of speech samples used for training and the differences between the individuals providing the samples. In applications where speech recognition models are to be used over a wide geographical region, it is particularly desirable that speech samples be collected from the various geographic regions where the models will ultimately be used. In this manner, regional speech differences can be taken into account during model training.
Another problem confronting implementers of speech recognition systems is that older speech recognition models may include different feature information than current speech recognition models. When updating a system to use newer speech recognition models, previously used models in addition to speech recognition software may have to be revised or replaced. This frequently requires speech samples to retrain and/or update the older models. Thus the problems of collecting training data and training speech recognition models discussed above are often encountered when updating existing speech recognition systems.
In systems using multiple speech recognition devices, speech model incompatibility may require the extraction of different speech features for different speech recognition devices when the devices are used to perform a speech recognition operation on the same speech segment. Accordingly, in some cases it is desirable to be able to supply the speech to be processed to multiple systems so that each system can perform its own feature extraction operation.
In view of the above discussion, it is apparent that there is a need for new and improved methods and apparatus relating to a wider range of speech recognition issues. For example, there is a need for improvements with regard to the collecting of speech samples for purposes of training speech recognition models. There is also a need for improved methods of providing users of portable devices with limited processing power, e.g., notebook computers and personal data assistants (PDAs) speech recognition functionality. Improved methods of providing speech recognition functionality in systems where different types of speech recognition models are used by different speech recognizers is also desirable. Enhanced methods and apparatus for updating speech recognition models are also desirable.